Amazon's New AI Transcribes Meetings and Names Who Said What

Amazon Web Services has published a guide and a ready-to-use package called WhisperX for its SageMaker AI cloud service. WhisperX is free, open-source software built on Whisper — OpenAI's speech-to-text model that turns audio into writing. Whisper by itself gives you a rough transcript. WhisperX adds two things ordinary transcription gets wrong: it stamps each individual word with a precise time, and it labels who was speaking. (The technical term for labeling speakers is "diarization" — think of it as name tags for voices.)

Why does that matter? Because a transcript without those details is hard to trust or use. Standard tools often place timestamps a few seconds off, which wrecks subtitles and redactions. And when nobody knows who said what, you can't run a compliance review or settle a legal dispute. With speaker labels and word-level timing, a recording becomes something you can search, clip, caption, and analyze at scale — the difference between a shoebox of cassette tapes and a neatly indexed filing cabinet.

The practical uses are everyday ones. Call centers can measure how long agents talk, check whether they followed a script, and spot unhappy customers automatically. Teams can turn meeting recordings into searchable notes. Media and online course makers can generate accurate captions for huge libraries. Hospitals, law firms, and banks can produce transcripts that stand up to audits. Output comes in formats video editors already use, so the same tool feeds both data dashboards and editing software. Short clips can be transcribed live; long recordings get submitted and picked up later.

The catch is setup and cost. It runs on Amazon's cloud, so you pay for computing time while it works — though it can shut down completely when idle. Real-time transcription is capped at roughly 60 seconds per request, so hour-long recordings go through the slower batch route. And like any cloud service, your audio leaves your building and lands on Amazon's servers, which matters if the recordings are sensitive.